2.1.2

What's Changed

• Fix NaN treatment in logsumexp (#369)
• Optimize segment_coo and segment_csr for torch.bfloat16 and torch.half (#375)
• Add PyTorch 2.1.0 support (#395)

New Contributors

• @jamesmyatt made their first contribution in #376
• @AgarwalSaurav made their first contribution in #377

Full Changelog: 2.1.1...2.1.2

2.1.1

• Added support for PyTorch 2.0

2.1.0

• Improved runtime of scatter_min, scatter_max and scatter_mul CUDA kernels (#336)
• Enabled ROCm build (#325)
• Added bf16 support (#316)

2.0.9

• PyTorch 1.10 support
• Allow loading of CPU wheels in a PyTorch CUDA installation #229 (thanks to @aphedges)
• Removed unnecessary eps argument in scatter_softmax #238 (thanks to @Veason-silverbullet)
• Resolved torch.floor_divide deprecation warnings #236 (thanks to @digital-idiot)
• Added a dim_size argument to scatter_softmax and scatter_log_softmax #243 (thanks to @mallamanis)

2.0.8

• Half-precision support (torch.half) for all operators in torch-scatter

2.0.7

• Prevent division by zero in scatter_std
• Fixed a bug in which an "index out of bounds" error occured in case of dim_size=None
• Fixed a bug in which absolute paths were added to SOURCES.txt

2.0.6

• Implementations of scatter_mul and segment_mul
• Reduced the size of shared library files
• CUDA wheels can now also operate on CPU-only devices

2.0.5

• Fixed a bug in which the backward pass of scatter_std did not succeed

2.0.4

Fixed a bug where composite functions could not be differentiated.

2.0.3

• Windows fixes
• Wheels for common OS/CUDA combinations